TR Forums

grantmeaname wrote:

qurious73ss wrote:

You think they have a different mask for Westmere and Gulftown? If you think they are different then Intel must throw away alot of die that don't meet the criteria for the top bin or maybe they use the same mask and can bin 5680X down to E5600 with i79xx in between. With record revenue and earnings, as well as margins >65%, my guess is that they still sell those chips that don't meet top bin criteria instead of throwing them away.

I think the fact that a business the size of Intel has a high margin doesn't not necessarily counterindicate a low-margin practice in a small halo-effect area of their business and that you're making a giant assumption in saying so.

A "giant assumption" because I assume that they don't throw away 90% of their die that don't meet the Extreme criteria? How hard is it to believe that since they are exactly the same chip that instead of throwing them away they bin out the wafer into several different segments including a client based extreme part.

qurious73ss wrote:

It works like this Westemere http://ark.intel.com/ProductCollection. ... Name=33174 & Gulftown http://ark.intel.com/ProductCollection. ... Name=29886. Same process and same wafer.

That's Westmere-EP, not Westmere. The former consists of Xeons which cost at least $850 for a 6-core CPU and usually a whole lot more than that (up to a factor of 2). Westmere is the die shrink of Lynnfield used for Arrandale (the laptop architecture) and such.

qurious73ss wrote:

grantmeaname wrote:

qurious73ss wrote:

Yes Gulftowns/Westmere share the ... same wafer

[Citation Needed]

You think they have a different mask for Westmere and Gulftown? If you think they are different then Intel must throw away alot of die that don't meet the criteria for the top bin or maybe they use the same mask and can bin 5680X down to E5600 with i79xx in between. With record revenue and earnings, as well as margins >65%, my guess is that they still sell those chips that don't meet top bin criteria instead of throwing them away.

Intel has big enough volume to do that with processors of different cores. Chip sizes for 4 vs 6 cores are not insignificant. Also since Nehalem Intel has been touting its "modular build" process meaning they can easily do varying cores, QPIs, and other building blocks. The Gulftown's are X56xx yes, but I am pretty certain that they don't share the same wafer with the 4-core chips.

AMD went with disabled cores for some of their processors because they simply did not have enough volume to justify the costs. However, the cost per chip is the same whether you are looking at the 3 or the 4 core models.

Althernai wrote:

qurious73ss wrote:

It works like this Westemere http://ark.intel.com/ProductCollection. ... Name=33174 & Gulftown http://ark.intel.com/ProductCollection. ... Name=29886. Same process and same wafer.

That's Westmere-EP, not Westmere. The former consists of Xeons which cost at least $850 for a 6-core CPU and usually a whole lot more than that (up to a factor of 2). Westmere is the die shrink of Lynnfield used for Arrandale (the laptop architecture) and such.

Yes, Westmere EP is the Xeon, I get confused with all the code names, but as I said Westmere EP and Gulftown are built on the same and wafer and this is how Intel bins their chips and is able to make money on the so called halo product.

Flying Fox wrote:

Intel has big enough volume to do that with processors of different cores. Chip sizes for 4 vs 6 cores are not insignificant. Also since Nehalem Intel has been touting its "modular build" process meaning they can easily do varying cores, QPIs, and other building blocks. The Gulftown's are X56xx yes, but I am pretty certain that they don't share the same wafer with the 4-core chips.

AMD went with disabled cores for some of their processors because they simply did not have enough volume to justify the costs. However, the cost per chip is the same whether you are looking at the 3 or the 4 core models.

Not sure about the 4 core Westemere EP either, but I too assume that the Gulftowns and X56XX are built on same wafer and as so is making money for them and not just a halo product. Here is some info I found regarding the manner.

http://techreport.com/discussions.x/18618
http://techreport.com/articles.x/18581

This is where I have trouble with your statements:

qurious73ss wrote:

Yes Gulftowns/Westmere share the same process and gasp! same wafer they are binned at a price range of $300-$1800. Again those $1800 chips cost as much as the $300 chips to manufacture and Intel makes alot of money selling them either as i7s or Xeons.

At that time you were confused about Westmere vs Westmere-EP. I just checked the Ark and I did not seem to find any 6-core Westmere-EP going for as low as $300. The ones that are being sold as Xeons should be put through a more stringent verification/testing process though, thus increasing the cost just a little bit.

Flying Fox wrote:

This is where I have trouble with your statements:

qurious73ss wrote:

Yes Gulftowns/Westmere share the same process and gasp! same wafer they are binned at a price range of $300-$1800. Again those $1800 chips cost as much as the $300 chips to manufacture and Intel makes alot of money selling them either as i7s or Xeons.

At that time you were confused about Westmere vs Westmere-EP. I just checked the Ark and I did not seem to find any 6-core Westmere-EP going for as low as $300. The ones that are being sold as Xeons should be put through a more stringent verification/testing process though, thus increasing the cost just a little bit.

Yes and yes. Those two products are very different and the cheapest6C Westmere EP is ~$900, the 4C are the $300 ones. So correction, they have actually binned these chips from $1700-$900. Does anyone still think their not making money on the i79XX chips?

I think that you've still not provided a single piece of evidence that they are. I could say, "Look, I found some sources, therefore Intel doesnot make a profit," and I would have provided just as much substantiation for my claim as you did.

And posting repeatedly in the same thread does not make your argument more convincing or intelligent.

grantmeaname wrote:

I think that you've still not provided a single piece of evidence that they are. I could say, "Look, I found some sources, therefore Intel doesnot make a profit," and I would have provided just as much substantiation for my claim as you did.

And posting repeatedly in the same thread does not make your argument more convincing or intelligent.

Now of course I cannot find a link that says they are making a profit on the i79XX but I do have some evidence that i79XX are low binned X5600 and if so then there is enough volume to make a profit. Actually this started as rebuttal at JFAMD who said that he believes that Intel is losing money making i79XX and that they are just a halo product since they're so low volume. I believe that they are actually low binned X5600 that are being sold as i79XX and provided a couple of links where it shows that the i79XX and X5600 are pretty much identical i.e. core, memory channels, die size, etc. http://ark.intel.com/ProductCollection. ... Name=33174 & http://ark.intel.com/ProductCollection. ... Name=29886. And here Scott Wasson says the same thing http://techreport.com/discussions.x/18618. Now comes the matter of volume, and maybe someone can dig up the numbers or has enough time to do the math but my honest opinion, which you probably don't care much for, is that their is enough volume in the server line to make a profit selling a $1K client part.

qurious73ss wrote:

Now of course I cannot find a link that says they are making a profit on the i79XX but I do have some evidence that i79XX are low binned X5600

Low-binned? The i7-900 chips both run north of 3GHz at speeds similar to the fastest Westmere-EP chips.

and if so then there is enough volume to make a profit.

Remember that assumption thing we talked about?

Actually this started as rebuttal at JFAMD who said that he believes that Intel is losing money making i79XX and that they are just a halo product since they're so low volume.

No, rebuttals have data. Moreover, your assertions have been modified a couple of times now.

Now comes the matter of volume, and maybe someone can dig up the numbers or has enough time to do the math but my honest opinion, which you probably don't care much for, is that their is enough volume in the server line to make a profit selling a $1K client part.

It's the highest bin!

grantmeaname wrote:

qurious73ss wrote:

Now of course I cannot find a link that says they are making a profit on the i79XX but I do have some evidence that i79XX are low binned X5600

Low-binned? The i7-900 chips both run north of 3GHz at speeds similar to the fastest Westmere-EP chips.

and if so then there is enough volume to make a profit.

Remember that assumption thing we talked about?

Actually this started as rebuttal at JFAMD who said that he believes that Intel is losing money making i79XX and that they are just a halo product since they're so low volume.

No, rebuttals have data. Moreover, your assertions have been modified a couple of times now.

Now comes the matter of volume, and maybe someone can dig up the numbers or has enough time to do the math but my honest opinion, which you probably don't care much for, is that their is enough volume in the server line to make a profit selling a $1K client part.

It's the highest bin!

So you do agree that they are both made on the same wafer, well that's a start. Now as for the speed of the chips, there is more to binning then speed, especially when it comes to server. I'm betting that the X5680 was binned a bit higher then the i7980 for a reason. As for how they can make money on these chips, well that just takes a little thought process

qurious73ss wrote:

So you do agree that they are both made on the same wafer, well that's a start.

Where did I say that? I believe you just made that up.

Now as for the speed of the chips, there is more to binning then speed, especially when it comes to server. I'm betting that the X5680 was binned a bit higher then the i7980 for a reason.

If only 5% of all chips can be sold as i7-980s, the fact that it's a lower bin than something else doesn't make them more economical than if they were a higher bin. The fact is, there's still only a fixed fraction of the chips that can perform that high, and putting more SKUs higher doesn't somehow increase that proportion.

As for how they can make money on these chips, well that just takes a little thought process

You're making an assumption, and a really stupid one at that. It doesn't just take a little thought, it takes an analysis of the economic realities involved. Wink.

qurious73ss wrote:

As you know those $1k chips and $200 chips come from the same wafer, so while you may only get 20 high bin chips out of 200 total chips it does NOT cost anymore to bin these chips. You do understand that for the most part Intel bins their chips out of the same wafer. There are a few exception of course such as Itanium, and Xeon-EX, but for the most part a Xeon and i7 are made with the same recipe. And I would hope that Amd's total revenue is greater then what Intel makes on the $1000 extreme chips, but the fact that Intel can sell $1k parts is why their margins are so high. Actually if you think about it, those $1K chips are really not the top bin of the wafers. Xeons come out of the same wafers too so they actually are able to sell some of those chips for $1800. Does Amd not bin there chips in the same manner?

Actually, you are right about the wafer, but also wrong. Take that $1000 core i7 part. I would be willing to bet that it does not occur in the standard distribution. On a typical wafter you get ~5% top bin and probably the other 95% are usually in N-1 and N-2 bins. Typically anything in N-3 or below is a downclock If you look at the prices of the intel parts you see:

3.33GHz = $1049
3.2GHz = $899
3.2GHz = $579
3.06GHz = $299

So here's my theory: top standard distribution is probably at the $579 price point. They are probably getting 5% yield to that speed with a high overall usable die ratio.

So, let's say for fun that they can theoretically get 200 die to the wafer (I have no idea what the actual is). If they run a standard wafer they get 5% $579 parts, probably 25% $299 parts and the rest is whatever is the next step down. But they are getting 200 good dies at the end of the day. So the cost to make anything from $579 down is (wafer cost) / 200.

But they sell so few $899 and $1049 parts, that to fill the demand they only need a handful of parts. So you run the process really hot in order to get those top bin speeds that don't occur naturally. You force the wafer to get you those really hot parts, but you end up with a ton of unsuable die. See, in the example above, you are pushing to maximum yields to make the most money possible. Here, you want speed, and you pay for it. You run the process hot and you get enough of those really high clock speed parts. But, your yield sucks and you only net out 30 decent die off the wafer. Now your cost is (wafer cost) / 30. Significantly higher price. Which is why it costs so much. If they made those parts $300 they would have too much demand and they would have a very unprofitable business. But if you price them high few will buy them, so they end up not having to make that many.

So, as you can see, even though it is all the same wafer mask, there are varying costs based on yield and distributions, you can't deny it.

When qurious realizes he's wrong he just stops responding.

Thanks for the explanation, though.

Chemically, how precisely does one run a 'hot' process? How is it that the results are sent to extremes?

Does it produce a bunch of chips that could be L3-cache-disabled garbage parts? That way I feel like they could at least break even on that two-hundredth of the diewafer...

Since I don't work in engineering I can's say specifically whether it is chemical or not (my guess is not).

Every wafer has a set of steps that take it from bare silicon wafer to finished product. You tweak the process to push it harder to run hot (more high speed parts, lower overall yields) or colder (higher yields, lower speed parts). Typically you "dial in" the process steps to get you a percentage yield of your top bin and then try to maximize on the die to ship ratio (how many working dies you can get out).

There are different things in the process that change, but the photomask set is always constant (that tells each step in the process what to lay down on the wafer) and the chemicals *should* be the same, but there might be different mixtures. Again, I have been through the fab on tours but I don't work with it day in and day out.

Each wafer that goes in has a recipe for how it is created and that recipe tells you what you should expect on the backend. You end up with a bunch of formulas and scenarios. Need more high bin speed, choose recipe A, need more parts, choose recipe B, need more low power parts, choose recipe C, and so on.

One of the more interesting jobs is in supply/demand where you look at the sales pipelines and determine what you need to fill them. Considering a wafer takes ~12-13 weeks on average to go from base to shipping parts, it's a dark science to call that correctly.

What you're saying makes a lot of sense. Based on how CMOS circuits work in general, I can think of at least a couple of potential tradeoffs: Tweaking the process to A) minimize transistor gate capacitance; and B) maximize transistor drive strength, will both improve speed. But the former will probably increase the number of defective transistors, and the latter will probably increase power consumption (which doesn't necessarily prevent the CPU from functioning, but may push more of the dice to the point where they don't fit into any of the defined bins).

I'm sure that's a gross oversimplification, but it probably captures the essence of the sort of issues that are involved.

JF-AMD wrote:

Every wafer has a set of steps that take it from bare silicon wafer to finished product. You tweak the process to push it harder to run hot (more high speed parts, lower overall yields) or colder (higher yields, lower speed parts). Typically you "dial in" the process steps to get you a percentage yield of your top bin and then try to maximize on the die to ship ratio (how many working dies you can get out).

Thanks a lot! I did not realize that the yields and performance could be tweaked or that there is an inverse relationship between them.

That said, I am not quite sure what the incentive to do this with the 980X is. When it came out (or even now), there were no consumer CPUs that come anywhere close to Gulftown in multi-threaded performance and a single multiplier step is not going to change that. Intel is effectively competing only with itself. Why would it run the process "hot" when it can get a lot more chips at a slightly lower frequency and sell them at the same price?

Althernai wrote:

That said, I am not quite sure what the incentive to do this with the 980X is. When it came out (or even now), there were no consumer CPUs that come anywhere close to Gulftown in multi-threaded performance and a single multiplier step is not going to change that. Intel is effectively competing only with itself. Why would it run the process "hot" when it can get a lot more chips at a slightly lower frequency and sell them at the same price?

Are you sure it's running hot?

Based on the price, they are definitely running it hot. If that was the naturally occurring top bin speed they would be selling that as top bin. They price it list that so that they don't have to build to many, probably too much scrap.

If you were a business analyst you'd take one look at their stack and conclude that the top two bins (at least) do not occur naturally without some "help". Because if they did occur in decent enough quantities, you'd price them at $400 and make a ton of revenue and margin because you'd sell 20-50X what they are selling at $999. The $999 price is s=designed to weed out all but a few. And get benchmarks.

Think about this, you are the product manager for that part. You can get 5% of your mix on a part that you can get $400 for. But instead you price it at $1000 and sell to .01% of the market. How long do you think you would have your job? Assuming that the cost on those is the same, and let's just make it $100 for fun (I have zero idea of what the cost is, this is only for example), you could sell 3 @ $400 and make as much as selling 1 @ $1000. Now, let's figure out the real math. If you can sell a million a month and 5% are at $400, that net revenue is $20M and the net margin is $15M. That $1000 part, at .01% of the market nets you $100K and $90K of profit. How does that pricing model cover the cost of business? It can't.

I think that tech report thread on price performance is the most accurate understanding that I have seen from the issue. If you take a look at where the curve goes completely haywire from a price/performance aspect you will see the "naturally yielding" parts and the unnatural parts.

grantmeaname wrote:

When qurious realizes he's wrong he just stops responding.

Thanks for the explanation, though.

Chemically, how precisely does one run a 'hot' process? How is it that the results are sent to extremes?

Does it produce a bunch of chips that could be L3-cache-disabled garbage parts? That way I feel like they could at least break even on that two-hundredth of the diewafer...

So do you still think the i7 9XX and Westemere EP fabd on different wafer?

JF-AMD wrote:

Since I don't work in engineering I can's say specifically whether it is chemical or not (my guess is not).

Every wafer has a set of steps that take it from bare silicon wafer to finished product. You tweak the process to push it harder to run hot (more high speed parts, lower overall yields) or colder (higher yields, lower speed parts). Typically you "dial in" the process steps to get you a percentage yield of your top bin and then try to maximize on the die to ship ratio (how many working dies you can get out).

There are different things in the process that change, but the photomask set is always constant (that tells each step in the process what to lay down on the wafer) and the chemicals *should* be the same, but there might be different mixtures. Again, I have been through the fab on tours but I don't work with it day in and day out.

Each wafer that goes in has a recipe for how it is created and that recipe tells you what you should expect on the backend. You end up with a bunch of formulas and scenarios. Need more high bin speed, choose recipe A, need more parts, choose recipe B, need more low power parts, choose recipe C, and so on.

One of the more interesting jobs is in supply/demand where you look at the sales pipelines and determine what you need to fill them. Considering a wafer takes ~12-13 weeks on average to go from base to shipping parts, it's a dark science to call that correctly.

Interesting theory. So if true then its even more impressive how Intel is still able to maintain ~%65 margins. They must be have discovered the secret formula to this hot/cold dial in process. BTW, in diffusion which is my forte, we have no such hot/cold recipes. We have tweaks for time, temp, press and set ones like gas ratios but a recipe is a recipe and its only tweaked to meet a control limit such as thickness of deposition but only when it falls out spec and that tweak is made for each specific tool and only certain things can be tweaked. What does affect the yield or characteristic of how well a wafer will yield is in our case of course is where it runs in the batch and as such will affect uniformity of the layer and of course keeping the wafer deposition or oxide layer as clean as possible but again no such thing as hot/cold recipe

qurious73ss wrote:

So do you still think the i7 9XX and Westemere EP fabd on different wafer?

I said you were making an assumption, which is true. Notice you're ignoring everything else that's been said in the thread...

grantmeaname wrote:

qurious73ss wrote:

So do you still think the i7 9XX and Westemere EP fabd on different wafer?

I said you were making an assumption, which is true. Notice you're ignoring everything else that's been said in the thread...

So you finally agree that they share the same wafer? As for assumptions, would you believe that I help make those wafers? Probably not

Let's look again at your initial claims and examine what I had an issue with and what I did not, and how these claims turned out.

qurious73ss wrote:

As you know those $1k chips and $200 chips come from the same wafer,

You later backtracked on this.

so while you may only get 20 high bin chips out of 200 total chips it does NOT cost anymore to bin these chips.

JF-AMD demonstrated why this was incorrect.

You do understand that for the most part Intel bins their chips out of the same wafer.

You later backtracked on this.

There are a few exception of course such as Itanium, and Xeon-EX, but for the most part a Xeon and i7 are made with the same recipe.

You later backtracked on this.

And I would hope that Amd's total revenue is greater then what Intel makes on the $1000 extreme chips,

You misconstrued what JF-AMD said; he never claimed this.

but the fact that Intel can sell $1k parts is why their margins are so high.

JF-AMD demonstrated why this was incorrect.

Actually if you think about it, those $1K chips are really not the top bin of the wafers. Xeons come out of the same wafers too so they actually are able to sell some of those chips for $1800.

This is irrelevant.

Does Amd not bin there chips in the same manner?

I'm so far lost as to what you're arguing here that I don't know how to respond.

In summary, for the six thousandth time, I never stated that Westmere-EP and Gulftown were not on the same wafer. I said that saying so was an assumption, and moreover, that statement was in reference to your statement that Westmere (non-EP) and the i7-980X were fabbed on the same wafer; this was further affected by the fact that it was a secondary point and the point of almost every one of my posts in this thread has been that you cannot make economic assumptions, because you're not so smart that you can extrapolate from one halo product in one division of a multibillion dollar company to the entire company's financial status, a statement I still stand by.

Did qurious73ss join TR just to troll this thread?

We all know JF-AMD and he makes smart informed comments and when he doesn't outright know something is a fact he explains that it is his opinion based on his particular experience.

qurious73ss, you have stated things as fact that are based on your opinion yet have not shared where you have come upon this group of opinions.

JF has explained things very well and even if I agreed with qurious73ss I'd still have a hard time figuring out what he's saying or even what his point is - other than trolling that is...

Just because I disagree with him doesn't make me a troll. He believes that Intel looses money on the Gulftowns and I believe that they make money on them.

JF-AMD wrote:

What you are taking into consideration is that the $1000 CPU and the $800 CPU result in high ASPs per CPU but extremely low total margin dollars because so few people buy them.

While few people buy i79XX, alot of companies and people buy Westmere-EPs and since both of these chips are manufactured on the same wafer they are really selling a Xeon in the client market. That should help with volume.

JF-AMD wrote:

Actually, you are right about the wafer, but also wrong. Take that $1000 core i7 part. I would be willing to bet that it does not occur in the standard distribution. On a typical wafter you get ~5% top bin and probably the other 95% are usually in N-1 and N-2 bins. Typically anything in N-3 or below is a downclock If you look at the prices of the intel parts you see:

3.33GHz = $1049
3.2GHz = $899
3.2GHz = $579
3.06GHz = $299

So here's my theory: top standard distribution is probably at the $579 price point. They are probably getting 5% yield to that speed with a high overall usable die ratio.

So, let's say for fun that they can theoretically get 200 die to the wafer (I have no idea what the actual is). If they run a standard wafer they get 5% $579 parts, probably 25% $299 parts and the rest is whatever is the next step down. But they are getting 200 good dies at the end of the day. So the cost to make anything from $579 down is (wafer cost) / 200.

But they sell so few $899 and $1049 parts, that to fill the demand they only need a handful of parts. So you run the process really hot in order to get those top bin speeds that don't occur naturally. You force the wafer to get you those really hot parts, but you end up with a ton of unsuable die. See, in the example above, you are pushing to maximum yields to make the most money possible. Here, you want speed, and you pay for it. You run the process hot and you get enough of those really high clock speed parts. But, your yield sucks and you only net out 30 decent die off the wafer. Now your cost is (wafer cost) / 30. Significantly higher price. Which is why it costs so much. If they made those parts $300 they would have too much demand and they would have a very unprofitable business. But if you price them high few will buy them, so they end up not having to make that many.

Now you don't have any evidence that they do and I don't have any evidence that they don't. Now with that being said, you would also think that this would have some impact on margins considering that Intel had record earnings and margins(~65%) last year and alot of that was due to their ~%90 server market share. So, could it possibly be that the top bin is their top bin and this so called "hot process" is not required. Also, a new top bin has just been released so why do that when you are already using a "hot process" and trying to keep demand down?

I hope Bulldozer does well simply so that Intel still has some sort of competition.

qurious73ss wrote:

:lol: Just because I disagree with him doesn't make me a troll.

No one is arguing that that is the case. Reading comprehension fail?

MethylONE called him a troll for the exact reason he mentioned.

I don't understand how 3.33 GHz could be an outsider binning when the i7s seem to generally be able to run much higher than that in overclocking on even stock voltage. I can't wait to see what Ivy Bridge can manage. I'm curious if 22nm is working out well.

swaaye wrote:

MethylONE called him a troll for the exact reason he mentioned.

I don't understand how 3.33 GHz could be an outsider binning when the i7s seem to generally be able to run much higher than that in overclocking on even stock voltage. I can't wait to see what Ivy Bridge can manage. I'm curious if 22nm is working out well.

Just because you can hit that speed in an overclock doesn't mean that it is naturally occurring in regular distributions. The whole idea of overclocking is that you are going beyond the limits that the manufacturer sets. If you pump me full of steroids I could probably hit a home run, but that doesn't mean I could do it without them.